Files
Neutron/docs/evidence/linux-session.md
T

178 lines
11 KiB
Markdown

# Local Linux fidelity capture
This joins the single final combined session described in
[windows-session.md](windows-session.md). Prepare locally; no early hands-on
session is requested. Every original/native acceptance obligation remains open
until actual evidence and human review. No gameplay/movement reset is added.
Use the existing sealed Linux package, not a rebuilt or relabeled package.
The 01-05 package is `artifacts/fidelity/01-05-packages/linux-x64`:
source `446930d2a30d6d616c60ac5e59e4169a840bea09`, dirty yes, publish diff
`90cef95669caece53fe2d2c4736cf2ac05ed27f5d25ca03ab572589686bcb73d`.
Its file-list hash is `124cf271e4952f3822eec5c24358447fcffc9e9e1803aded8d96d4d4472dc5a3`.
The collector checks all sealed bytes before launch and collection. Keep output
outside the package and originals, in a fresh session directory.
Select the actual local X11 display and capture rectangle, real Pulse monitor
source and physical input method. Save a local JSON settings object with these
exact fields (values below are examples, not observed devices):
```json
{"display_method":"x11","display":":0","audio_method":"pulse",
"audio_source":"actual-device.monitor","input_method":"physical",
"size":"1280x720","offset":"0,0"}
```
Inspect `pactl list short sources` and `pactl list short sinks` locally. Null/dummy
devices do not qualify. This collector implements X11/Pulse capture with installed
FFmpeg/ffprobe; it records unsupported Wayland/TTY/dependency conditions as failed
preflight. Do not install or select a guessed alternative automatically. Use a
separately identified existing recorder for unsupported display sources and retain
its actual settings, tool identity, raw recording and timing provenance.
```sh
bash scripts/collect-fidelity-linux.sh preflight --package artifacts/fidelity/01-05-packages/linux-x64 --settings /local/linux-settings.json
bash scripts/collect-fidelity-linux.sh run --package artifacts/fidelity/01-05-packages/linux-x64 --settings /local/linux-settings.json --output artifacts/fidelity/linux/unique-launch --session-id unique-launch --case-id fresh-natural-ordinary --scope dialogue-return --seconds 120
bash scripts/collect-fidelity-linux.sh collect --run artifacts/fidelity/linux/unique-launch --import-to artifacts/fidelity/linux-receipts
```
Nonzero run/collect exits preserve failure/abort output; inspect it, and collect
the failed run too. Recorder starts before the normal `--load-save NewGame.tsk
--walk` process and passive observer. This is the existing story entry, not an
implemented main menu. Close the game normally before the selected duration.
Both children have one bounded deadline; interruption ends only owned children.
Logs and metadata have explicit bounds; raw recording is capped at 8 GiB. Retain
the entire lossless FFV1/PCM MKV and tail; a detected frozen interval is an open
inspection flag, not a claim of failure or parity derived from a numeric tolerance.
Collection probes actual streams/duration and hashes logs/media/observer and
package/source/resource identities. The duplicate MKV audio alias preserves exact
bytes for the shared receipt's distinct audio role. No audio normalization,
retiming, crop repair or preview is implicit. If a browser needs a remux/preview,
retain raw media and explicit derivation command/version/hash/time mapping under
the existing shared derivative contract. The local viewer exporter consumes the
imported receipts; actual playback/seeking/audio/crop remain human review gates.
Record manual event annotations as `{"events":[...]}` using the exact fields in
`scenarios.json`: hashed raw artifact name, media and monotonic elapsed seconds,
method and uncertainty. Supply `collect --annotations /local/annotations.json`.
Port simulation ticks remain separate from raw clocks. Observer gate results
do not prove a physical key was pressed. Automated injection is `port_fixture`;
test doubles are permanently synthetic. No collector output grants acceptance.
## Actual 2026-11-13 preflight
The scoped host probe, outside sandbox device restrictions, found a TTY session,
no `DISPLAY` or `WAYLAND_DISPLAY` and no X11 socket. Existing FFmpeg/ffprobe are
available. Pulse exposes only `auto_null` and `auto_null.monitor` (suspended).
No Xvfb or OBS command resolved. A synthetic display or null audio would not
establish the requested normal visible/audible gameplay, so no game/recorder
was launched and no substitute route/media was fabricated.
The exact failed focused receipt is `linux-01-06-focused-preflight-linux`, at
`artifacts/fidelity/linux/01-06-focused-preflight/receipt.json`, imported into
`artifacts/fidelity/linux-receipts/linux-01-06-focused-preflight-linux/`.
`preflight.json` preserves observed OS, tool versions/hashes, Pulse devices and
available GPU/Vulkan diagnostics; missing SDL/active-instance runtime versions
remain explicit. The selected package/resource seal matches the hash above.
Both the failed result and its import validate structurally, with acceptance open.
Before collection, all 539 original files verified unchanged.
Required next physical session: provide an actual display/game capture source
and real audible device/monitor source, verify moving unobstructed content and
audio, then record physical normal controls and full tails. Inspect capture cost,
crop/aspect and actual browser playback/seeking separately. Preserve this failed
preflight as history; new attempts use new session IDs and fresh output roots.
## Full matrix and the final combined agenda
The current typed dependency receipt is
`linux-01-06-matrix-preflight-linux`, at
`artifacts/fidelity/linux/01-06-matrix-preflight/receipt.json`, imported under
`artifacts/fidelity/linux-current-receipts/`. The earlier task-1 receipt is
retained diagnostic history; select the current typed receipt for scenario audit.
`artifacts/fidelity/linux/01-06-matrix.json` enumerates every obligation, exact
procedures/events, failed attempt ID and shared display/audio dependency.
There were no runnable native cells after that common dependency failed. The
collector does not manufacture 40 failed game launches or unrelated scene clips.
In this table **OPEN** means the exact current dependency receipt above plus
the named cell/obligation's actual recording, lifetime evidence and human review.
| Cell | continuous | dialogue-return | bank-left | bank-right | landing-idle |
|---|---|---|---|---|---|
| fresh-natural-edge | OPEN | OPEN | OPEN | OPEN | OPEN |
| fresh-natural-ordinary | OPEN | OPEN | OPEN | OPEN | OPEN |
| fresh-skipped-edge | OPEN | OPEN | OPEN | OPEN | OPEN |
| fresh-skipped-ordinary | OPEN | OPEN | OPEN | OPEN | OPEN |
| retained-natural-edge | OPEN | OPEN | OPEN | OPEN | OPEN |
| retained-natural-ordinary | OPEN | OPEN | OPEN | OPEN | OPEN |
| retained-skipped-edge | OPEN | OPEN | OPEN | OPEN | OPEN |
| retained-skipped-ordinary | OPEN | OPEN | OPEN | OPEN | OPEN |
Follow `scenarios.json` literally. Continuous recording begins before startup or
retained setup and keeps lab → VR04 → lab → outside in one process/clock and raw
video. Use normal supported controls/gadget/RETURN, without warps/grants. Focused
clips have their own declared setup and never substitute for the continuous run.
Natural dialogue reaches its real end; skipped dialogue records the actual skip
key and observed skip/end boundaries. Ordinary play releases movement at return;
edge play records held input through dialogue and RETURN, release/first-neutral,
independent left/right banked jumps, landing and idle. Preserve legitimate bank,
TSK-retained story input order and distinct fresh direct-scene observations.
No blanket angle/pixel/timing/audio tolerance or leveling repair is implied.
For each fresh observation, use a new launch/session ID and separately identified
generated execution copy; verify its existing seal rather than rebuilding it.
Keep the source package immutable. The collector records the selected package
root and sealed pre/post identities; copying/staging is an explicit local step,
not an invented copy claim. The prepared checkpoint's immutable Linux package is
also available for a verified local copy. Never select the original installation.
For retained observations, keep one actual process, copy and clock across visible
prerequisite lab departure to VR04 and return, then the tested interval. Capture
the real generated-copy state while that process remains alive; a later launch
or synthetic snapshot cannot replace the prerequisite. Use a separate terminal
to save the observed state records inside the run directory before finalization.
Annotations may declare `artifact_roles` mapping contained recording paths to
`prerequisite`, `identity_files` mapping `copy_pre`, `retained_copy`, `copy_post`
to actual contained state-record paths, and `continuity` with `lifetime_id`,
`prerequisite_artifact`, `copy_pre_sha256`, and for retained runs
`departure_session_id`, `return_session_id`, `retained_copy_sha256`.
All hashes refer to actual identity records. The selected prerequisite recording
must have a probed duration, and prerequisite events precede the tested interval.
A byte-identical alias of raw media can identify its prerequisite role without
retiming; preserve its exact bytes. The collector never invents these observations.
`clocks.clock_id` identifies the outer capture clock; preserve the passive observer
clock's separate origin and annotate any alignment/uncertainty explicitly.
Export/update the checklist with an unused output filename:
```sh
bash scripts/collect-fidelity-linux.sh matrix --receipt artifacts/fidelity/linux-current-receipts/linux-01-06-matrix-preflight-linux/receipt.json --dependency-run artifacts/fidelity/linux/01-06-matrix-preflight --output artifacts/fidelity/linux/new-matrix.json
```
For interrupted work, pass `--previous /local/prior-matrix.json` and repeat
`--receipt` for new attempts. Old receipts are re-read and hash-checked; reused
or conflicting session/receipt IDs are rejected. Preserve failed/incomplete
history, retry only missing/incomplete obligations, and use fresh IDs/directories.
Do not hand-edit history to imply completed cells. Shared scenario validation
checks chronological required markers, one recording and retained identity;
the final acceptance audit still owns review and applicability selection.
The single final session adds these Linux actions to the Windows/original agenda:
1. Verify all 539 authoritative originals before staging and after collection with
`python3 tools/recon/inventory.py ../Original --verify docs/original-manifest.json`.
2. Select real display/audio and identified installed recorder/probe; perform a
short moving-content, audible-output, crop/aspect and physical-control check.
3. Record all 40 obligations with their real setup/lifetime; retain raw video/audio,
accepted gate/observer logs, manual input descriptions, clocks and uncertainty.
4. Import current applicable receipts and export local paired review. Verify actual
browser playback/seeking/tails at normal speed and exclusive original/port audio;
retain any preview's exact derivation and raw source. Review capture overhead.
5. Recheck source/package/original identities and matrix history, retain remaining
discrepancies/open dependencies, and transfer results locally together once.
The Linux subset is **0/40 accepted**. The full three-side protocol remains
**0/120 accepted**: original-DxWnd/native Windows/physical audible review are open.